Shuttlecock Hitting Event Detection & Rally Analysis in Badminton (CV + ML)

This repository contains the implementation for a research project on shuttlecock hitting event detection and rally analysis in badminton using computer vision and machine learning.
The pipeline extracts hit events from match videos, constructs hit-windows to handle timing uncertainty, engineers baseline (F0) and enhanced (F1) rally features, and evaluates winner prediction models using leak-safe GroupKFold splitting by video.

---

Project Overview

Goal: Automatically analyze badminton match videos by:

1. Detecting shuttlecock hit events
2. Segmenting rallies and extracting rally indicators
3. Engineering match-level feature sets (F0 and F1)
4. Training/evaluating ML models to predict the match winner

---

Key Contributions

• Hit event → Hit-window sampling: Constructs temporal windows around predicted hit frames (offset k ∈ {0, 3, 5}) to increase robustness under small timing shifts.
• Rally analytics pipeline: Converts hit events into rally-level indicators (tempo, rally length, inferred zones/landing proxies, etc.).
• Feature engineering:
  • F0 (Baseline): fundamental hit/rally statistics
  • F1 (Enhanced): adds domain-informed spatial/tempo/distribution features
• Winner prediction: Compares models (e.g., Logistic Regression / Random Forest / SVM / XGBoost*) and feature sets (F0 vs F1 vs F0+F1).
• Leak-safe evaluation: Uses GroupKFold by VideoName to prevent train/test leakage across the same match video.

* XGBoost is optional depending on environment.

---

Results Highlights (Update with your final numbers)

• Hit-window shuttle visibility:
  • k=0: 39.62%
  • k=±3: 64.90%
  • k=±5: 74.16%
• Winner prediction:
  • Best model: [Fill in: e.g., Random Forest + F1]
  • Accuracy / Macro-F1: [Fill in your final mean scores]
  • Tuning improved RF accuracy (example): 0.6816 → 0.6916 (update if needed)

---

Repository Structure